AREPO: Uncertainty-Aware Robot Ensemble Learning Under Extreme Partial Observability

Abstract

Real-world applications of vision-based robot learning face two major challenges: extreme partial observability and effective simulation-to-reality (sim-to-real) transfer. This letter introduces a robust robot learning framework that enhances uncertainty awareness to address these challenges. We reinterpret variational-autoencoder–based visual reinforcement learning (RL) from an uncertainty-quantification perspective, enabling resilience to high sensory noise and severe visual occlusions—common in industrial robotic tasks. To further improve sim-to-real transfer, we propose an uncertainty-aware ensemble RL algorithm. We validate our methods on a laboratory task designed as a proxy for real-world industrial applications characterized by harsh environments with low visibility and physical occlusions. Both simulation and real-world results demonstrate significant improvements in task accuracy and efficiency over various baselines, highlighting the benefits of uncertainty-aware robot learning for complex operational contexts.